How to reduce bias in LLM-as-a-judge evaluations

The symptom: the verdict changes when the evidence does not

You ask a model to pick the better of two answers, grade a draft or score a batch of resumes, and the result depends on things that have nothing to do with quality: which option came first, whose style the text resembles, what the previous score was, how confidently the text praises itself, or whether a candidate is labelled as yours. The scores look reasonable one by one. The bias only shows up when you change something irrelevant and the verdict moves.

The size of the effect is not small. In an audit of image-edit judges, swapping option A and option B flipped up to 60.9% of verdicts for one model, while repeating the same request in the same order changed only 1–10%. A judge that sees a previous score in the prompt lowered its new rating in 7 out of 8 models tested, even though the text was identical.

Why LLM judges are biased

A judge model does not measure quality in isolation. It reads the whole context and continues it, so every cue in the prompt is treated as evidence: position, labels, earlier verdicts in the chat, the wording of the answer under review. Four mechanisms recur across the studies:

• Order effects. Coding judges in CodeJudgeBench tended to prefer the option that came last; another position-bias test notes that primacy or recency can win and which one is hard to predict.
• Style recognition. Judges favour text that looks like their own. In a study of 20 models, self-preference was stronger in the more capable judges, not weaker.
• Context momentum. A run of rejections in the same chat makes the next rejection more likely; across 11 models, accumulated history hurt 1.6 times more than a run of approvals helped and reached its full size after 5 previous ratings.
• Coarse scales. Asked for 0–100, a GPT-5 judge used only about 20 distinct values, so 67% of comparisons ended in a tie and the choice became a coin toss.

Fixes ranked by measured effect

1. Let the judge answer first, then compare. For checks with a correct answer, de-anchoring has the model write its own answer before it sees the one under review; false positives fell from 72% to 1.2% in that study.
2. Compare pairs instead of scoring. Replacing 0–100 scores with "A, B or TIE" cut ties from 60% to 4% and raised the value of the choice from 21% to 61% on Chatbot Arena prompts.
3. Run both orders and keep only agreeing verdicts. Swapping positions exposes order bias that a plain repeat cannot; if the winner changes with the order, treat the pair as a tie.
4. Remove scoring history. No attempt numbers, no "previous score". With that metadata, GPT-4.1's probability of the top score dropped from 93% to 0%. For batches, give each item a fresh chat.
5. Split one verdict into criteria. Judging relevance, accuracy, depth, logic and clarity separately and summing the wins reduced self-preference by 31.5% on average.
6. Write the evaluation plan before judging. Declaring primary and secondary criteria, with length marked as secondary, reduced length bias by 16-32% in reasoning models and 7-17% in ordinary LLMs.
7. Narrow the criterion and average several samples. One sentence naming what to focus on plus eight independent ratings added 11.9 pp of accuracy together.
8. Strip cues that are not the work. Replace "ours/theirs" with neutral labels: across six model families, group-label bias disappeared without labels. Remove raw data previews that make a judge invent rules, and demand evidence so confident wording does not count as merit.
9. Pick the answer format deliberately. When many options are listed, Llama-3.1-8B chose the last option 20 times more often than the first; asking for a probability per option sidesteps that. For binary calls, Claude showed a yes/no ordering bias of up to −0.86 that a 0–10 scale avoids.
10. Tune the rubric on your own disagreements. Reflect-and-revise turns a vague rubric into explicit thresholds over 3-10 rounds of comparing the judge's scores with yours.

A combined prompt stack

For a decision that matters, layer the fixes rather than picking one:

1. Clean the input: anonymous labels (Text A, Text B), no authorship, no prior scores, no other people's opinions.
2. State one primary criterion and, if needed, separate sub-criteria, with length and polish named as secondary.
3. Ask for reasoning per criterion, then a verdict of A, B or TIE, then a confidence rating.
4. Run it in a new chat with the order reversed. Accept the winner only if both runs agree.
5. If there is a correct answer, add a step zero where the judge writes its own answer before seeing the candidates.

The prompts below are the building blocks for these layers. They cost more calls than a single score, but every layer targets a bias that was measured, not assumed.

What does not work

• Telling the judge to be unbiased. The standard MT-Bench prompt already says to avoid positional bias and ignore length, and it was the prompt with the highest self-preference.
• "Think step by step" or "ignore the previous score". Neither removed anchoring on numeric scores; the explicit warning helped only for right/wrong decisions.
• "Evaluate each item independently" inside one long chat. It reduces history drift but does not remove it; only a fresh context does.
• Calibrating numeric scores after the fact. It recovered only +2.6% in the best-of-N study; calibration examples and adaptive routing gave no reliable gain in the ensemble study.
• Asking more models. Judges disagree with each other almost completely (agreement 0.04 out of 1.0), and for correctness checks they share the same blind spot. A model grading its own output is the weakest setup of all: it approves its own bugs.

One honest contradiction: a grading study found a 0–5 scale agrees with humans better than 0–10, while the pairwise study argues against absolute scales altogether. Both can hold: if you must score, use a short anchored scale; if you must choose, compare pairs.

Evaluate two versions of {what_we_compare}.

Version A:
{version_A}

Version B:
{version_B}

Selection criterion: {criterion}

First explain how they differ on this criterion.
Then give your verdict: A / B / TIE.
Rate your confidence: 1 (nearly equal) — 5 (clear winner).

Evaluate two versions of {content type}.

Text A:
{first version}

Text B:
{second version}

Rate EACH criterion separately. For each one: the winner (A / B / tie)
+ one sentence on why.

1. {criterion_1} — {what exactly we're evaluating}
2. {criterion_2} — {what exactly we're evaluating}
3. {criterion_3} — {what exactly we're evaluating}
4. {criterion_4} — {what exactly we're evaluating}
5. {criterion_5} — {what exactly we're evaluating}

Summary: add up the wins per criterion and name the overall winner.

You are a {expert role} evaluating {item type}.

This is the only item you are seeing.
Evaluate it strictly on its own content.

Evaluation criteria:
— {criterion 1}
— {criterion 2}
— {criterion 3}

{item to evaluate}

Output: {answer format: PASS/FAIL or YES/NO}
+ {number} specific reasons.